Method and system for determining text coherence

1. A computer-implemented method of evaluating the coherence of an essay, the method comprising: receiving an essay with a computer, wherein the essay comprises one or more discourse elements and a plurality of text segments; annotating the one or more discourse elements of the essay; generating a text segment vector for each text segment corresponding to a discourse element using a vector-based method of random indexing with the computer; identifying a plurality of essay dimensions with the computer, each representing relatedness of a text segment to an aspect of the essay, each dimension having an associated dimension score that provides a measure of said relatedness, the essay dimensions including a first dimension measuring relatedness between a first text segment corresponding to a first discourse element and a second text segment corresponding to a second discourse element, and a second dimension measuring relatedness between text segments within a discourse element; measuring semantic similarity with respect to at least one text segment based on the essay dimensions using a support vector machine to calculate a first dimension score associated with the first dimension and a second dimension score associated with the second dimension; and generating one or more values representing a coherence level to the essay based on the first dimension score and the second dimension score with respect to the at least one text segment.

2. The method of claim 1 wherein the data provided to the support vector machine comprises one or more of the following: a maximum prompt similarity score for a text segment with a sentence in a prompt; a task sentence similarity score for the text segment with a required task sentence, wherein the required task sentence is a portion of the prompt including an explicit directive to write about a specific topic; a maximum thesis similarity score for the text segment with a sentence in a thesis; a maximum similarity score for the text segment with a sentence in a preceding discourse element; a predetermined text similarity score for the text segment with each of one or more predetermined text segments; a number of sentences in a discourse element corresponding to the text segment; a number of sentences in a discourse element corresponding to the text segment having a prompt similarity score greater than a first specified threshold; a number of sentences in a discourse element corresponding to the text segment having a task sentence similarity score greater than a second specified threshold; a number of sentences in a discourse element corresponding to the text segment having a thesis similarity score greater than a third specified threshold; a length, in words, of the text segment; a Boolean flag indicating whether the text segment contains an anaphoric element; a discourse element corresponding to the text segment; a thesis sentence; and a sentence numbering position, wherein the sentence numbering position corresponds to a number of sentences that the text segment is from the beginning of the discourse element corresponding to the text segment.

3. The method of claim 2 wherein the maximum prompt similarity score is computed by computing a prompt similarity score for each sentence of the prompt and the text segment and selecting a prompt similarity score that is greater than all other prompt similarity scores.

4. The method of claim 2 wherein the maximum thesis similarity score is computed by computing a thesis similarity score for each sentence of the thesis and the text segment and selecting a thesis similarity score that is greater than all other thesis similarity scores.

5. The method of claim 1 wherein the discourse elements comprise one or more of: background material; and a supporting idea.

6. The method of claim 1 wherein a text segment is assigned a rank based on its relatedness to the thesis.

7. The method of claim 1 wherein the discourse elements comprise a main idea.

8. The method of claim 7 wherein a text segment is assigned a rank based on its relatedness to the main idea.

9. The method of claim 1 wherein annotating comprises annotating each non-thesis discourse element by a human evaluator.

10. The method of claim 1 wherein annotating comprises annotating each non-thesis discourse element by an automatic evaluator.

11. The method of claim 1 wherein the plurality of essay dimensions further comprises a third dimension measuring relatedness of a text segment to a prompt, wherein the essay is written in response to the prompt; wherein the one or more values representing a coherence level to the essay is generated based on a third dimension score associated with the third dimension.

12. The method of claim 11 wherein the plurality of essay dimensions further comprises a fourth dimension measuring the number of errors in one or more of grammar, usage and mechanics for a text segment; wherein the one or more values representing a coherence level to the essay is generated based on a fourth dimension score associated with the fourth dimension.